Foundation Model
A large AI model trained on broad, diverse data that can be adapted to a wide range of downstream tasks through fine-tuning or prompting. The backbone of modern generative AI.
Foundation models represent a paradigm shift from task-specific models to general-purpose models that serve as a base for many applications. The term was popularized by the Stanford Center for Research on Foundation Models.
Key characteristics:
- Scale: Trained on massive datasets with billions of parameters
- Adaptability: A single foundation model powers many applications through fine-tuning
- Emergent Capabilities: Abilities that appear at scale but aren’t explicitly programmed
- Pre-training + Fine-tuning: Pre-trained once, then specialized for specific use cases
Examples include GPT-4 (OpenAI), Claude (Anthropic), Gemini (Google), LLaMA (Meta), and DALL-E (OpenAI). Foundation models can be closed-source (GPT-4) or open-source (LLaMA, Stable Diffusion). They form the core of most modern AI applications, from chatbots to code assistants to creative tools.